Anthropic published The components of agent memory on November 6, 2026, describing how an agent can reorganize experience between sessions. The research examines a process often called dreaming, in which the system reflects on transcripts to correct mistakes, produce notes and build more useful memory representations.
For teams deploying assistants over repeated tasks, the relevant question is what becomes easier or more reliable in the next session. A larger store of past conversations alone does not establish that an assistant will retrieve the right information or recognize when something it remembers needs correction.
Memory involves several different representations
Anthropic describes a hierarchy that includes original transcripts, notes, more structured representations such as wikis and context loaded into a new session. Each layer serves a different purpose. A transcript preserves the original interaction, while a short note can make a relevant fact easier to find later.
The research uses memory to mean knowledge retained from experience across sessions. It distinguishes this from changing model weights and from compressing a long conversation so that it can continue. That vocabulary helps teams describe what a persistent assistant is actually doing.
What idle-time review can add
The earlier Sleep-time Compute research from Letta and UC Berkeley provides a related foundation for using time between interactions to reorganize context. Anthropic’s work explores how that idea can help an agent improve the material it will rely on later.
Our practical interpretation is that useful memory requires a correction path. If a user changes a deadline, the assistant should be able to update the relevant representation and explain which current record it is using. Keeping the original transcript also makes it possible to inspect whether the correction accurately reflects what the user said.
Evaluate memory with changes, not only repetition
A meaningful test can begin with a small assignment that depends on facts learned in an earlier session. After the first attempt, change one fact and introduce a conflicting outdated reference. Check whether the next session uses the current information and whether the agent can recover its provenance.
Also test an intentional removal. A system that stores useful preferences should have a reliable way to stop using them when a person asks. That can be assessed at the output boundary by repeating a task that previously depended on the preference and inspecting what context the assistant retrieves.
The product implication
The research is relevant to recurring analysis, repository maintenance and assistants that revisit an evolving project. In each case, the benefit depends on selecting information that remains useful and preserving a way to challenge it. Memory quality can be assessed through task outcomes and correction behavior.
A team considering this approach can start with explicit notes and a small transcript set, then compare repeated-task performance with and without a review pass. Record incorrect carryover, retrieval misses and time spent correcting the assistant. This creates evidence about whether the memory process improves the work it is intended to support.